Highly accurate protein structure prediction for the human proteome
Level 5 - mechanism / opinion, no new human data
Computational structural modeling and bench bioinformatics analysis (graded Level 5 by design analogy, not clinical CEBM).
OpenAlex W3183475563 · doi:10.1038/s41586-021-03828-1
What was done
The authors applied the AlphaFold machine learning model across 98.5% of the human proteome to generate computational structural models. They introduced interpretation metrics built on AlphaFold outputs to assess multi-domain configurations and detect disordered regions, presenting case studies to demonstrate biological hypothesis generation.
What was found
Compared to experimental structures covering 17% of total human protein residues, AlphaFold generated models covering 98.5% of human proteins. Confident predictions were achieved for 58% of all residues, with 36% of all residues reaching very high confidence.
Why it matters
This resource drastically expands structural coverage of the human proteome, providing an open-access reference to accelerate mechanistic biology and structure-guided drug discovery.
Limits
The structures are computational predictions rather than experimentally resolved coordinates. Overall, 42% of residues lacked confident predictions, and the abstract does not report data on protein complexes, dynamics, alternative conformations, or post-translational modifications.
Cited by
- supports Demis Hassabis and his team solved the protein folding problem computationally, enabling 3D structural modeling of thousands of proteins.